Professional Documents
Culture Documents
Multinomial and Conditional Logit Models
Multinomial and Conditional Logit Models
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=paa.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Population Association of America is collaborating with JSTOR to digitize, preserve and extend access to
Demography.
http://www.jstor.org
Demiiograplhy,
Vol. 25, No. 3, August
1988
Multinomialand Conditional
LogitDiscrete-Choice
Modelsin Demography
Saul D. Hoffmnan
Departmentof Economics,University
of
Delaware,Newark,Delaware 19716
GregJ.Duncan
Institute of
forSocial Research,University
Miclhigan,Ann Arbor,Michigan48106
Statistical
techniques forthe analysisof discretechoices have beeinused withincreasinlg
regularityin demographic analyses.I The best known are the binomial logit and probit
techniques,bothofwhichare suitableforbinarychoiceproblems.For problemsinvolving
thechoice amongthreeor morecategories, themultinomial logittechniqueis mostoften
employed;the corresponding probitmodelis used relativelylittlebecause of itscomputa-
tionaldifficulty.Virtuallyunusedthusfaris a closelyrelatedtechniquecalled conditional
logit,a modelthatis wellsuitedforbehavioralmodelingofpolychotomous choicesituations.
DevelopedbyMcFadden(1973),theconditional logitmodelis widelyusedin transportation
demandstudies(see Ben-Akivaand Lerman,1985) but is seldomused in demographic
research.2
Conditionallogit is not simplya different and arguablypreferable techniquefor
estimating thekindof modelsforwhichmultinomial logitis currentlyused. Rather,it is
appropriate classofmodelsin whicha choiceamongalternatives
fora different is treatedas
a functionof the characteristicsof the alternatives,ratherthan (or in additionto) the
characteristicsoftheindividualmakingthechoice.
We believethatmanyproblemsofinterest to demographers and othersocialscientists
can be modeledby usinga "characteristics of the alternative"approach.Thus theyare
appropriately estimatedwithconditionial logit. Furthermore, we suggestthatit is often
toattacha behavioralinterpretation
difficult totheresults ofmodelsthatfocusexclusively on
the"characteristicsof thechooser"-thatis, thoseestimated byconventional multinomial
logit.
The nextsectionofthisarticledescribes thebasicstatistical
propertiesoftheconditional
logit(CLGT) modeland comparesit withthebetterknownmultinomiallogit(MNLGT)
model.3 It also considers
theformoftheunderlyin-g modelsofindividualbehaviorthatlead
(D 1988Populationi
Copyright Associationi
ofAmicrica
415
416 Vol. 25,No. 3, August1988
Demography,
Statisticaland ModelingIssues
Both multinomiallogitand conditionallogitare used to analyzethe choice of an
individualamonga setofJalternatives. The centraldistinctionbetweenthetwo,canbe put
verysimply:MNLGT focuseson the individualas the unit of analysisand uses the
individual's as explanatory
characteristics CLGT focuseson thesetof
variables;in contrast,
alternativesforeach individualand the explanatory variablesare characteristics
of those
alternatives.
4
ofthejth
ofindividuali andZi,forthecharacteristics
LetXi standforthecharacteristics
forindividuali, withthecorrespondinig
alternative parameter vectorsdenotedby/3and a,
respectively.Let Jbe the numberof unorderedalternatives (forthe moment,assumed
constantforall individuals) thatindividuali choosesalternative
andPi,theprobability j. The
choiceprobabilitiesin theMNLGT and CLGT modelsare
J
MNLGT: P11= exp(Xi/31) I exp(Xi/3k), (1)
Jk=
ofthealternatives
In a mixedmodelthatincludesbothcharacteristics and theinidividual,
the
corresponding canlbe written
probability as
J
Mixed: Pj; = I exp(Xifjl+ Z1jfl)/Cxp(Xi,Pk
+ Zika). (3)
k= I
vil
=
f(Ziv), (6)
StatisticalProperties
and EstimationIssues
In thissection,we presenta briefsurveyof some of the practicalstatistical
issues
involvedin theestimationoftheMNLGT and CLGT models.Amongotherthings,we call
attentionhereto one ofthepotentiallyundesirable imposedbythelogisticform
restrictions
used in eithermodel. We also discusssome estimation
issues.
LikelihoodFunction. Despite the differences
discussedearlier,the CLGT and
MNLGT modelssharea commonlikelihoodfunction:
log L = ' (ll)
yi,P
Dependent
Alternativevariable Z XD2 XD3
Individual
1 1 0 Z,1 0 0
2 0 Z12X1 0
3 1 Z13 0 Xi
2 1 0 Z21 0 0
2 1 Z22 X2 0
3 0 Z23 0 X2
3 1 1 Z31 0 0
2 0 Z32 X3 0
3 0 Z33 0 X3
4 1 0 Z41 0 0
2 1 Z42 X4 0
3 0 Z43 0 X4
Multinomialand ConditionalLogitModels 421
An EmpiricalExample:Remarriageand WelfareChoices of
Divorcedand SeparatedWomen
In thissection,we examinetheremarriage and welfarechoicesofdivorcedorseparated
women.We presentestimates ofthreemodels-a standard MNLGT modelwithindividual
characteristics as explanatory variables,a pure CLGT model withcharacteristics of the
alternativesas explanatory variables,and a mixedmodelthatincludescharacteristics ofboth
theindividualand thealternatives.
Our analysisis based onldata fromthe Panel Studyof IncomeDynamics(PSID) on
whitewomenunderthe age of 45 who becamedivorcedor separatedbetween1969 and
1982. Each womanis observedfromthedateofherdivorceor separation untilremarriage,
theend ofthepanelobservation period,orthesixthpost-divorce/separation year,whichever
comesfirst. Our dataare in person-year event-history format, definedovera spellofbeing
"unmarried."Formally,we are estimating a discrete-time hazard model of time until
remarriage, usinlg an MNLGT or CLGT modelas theestimation procedure.Time-varying
independent variablesare measuredas oftheperson-year used in theanalysis.See Allison
(1982, 1984) fora generaldiscussionofdiscrete-time hazardmodels.1
In each year,a womanis observedin one ofthreealternative states:she can remarry,
she can remainsingleand receivewelfare,or she can remainsinglewithoutreceiving
welfare.We use functionial ratherthan legal definitionsof marriage,divorce,and
remarriage. Unmarried couplesaretreated as married bythePSID iftheyresidetogether for
twoconsecutive interviews;giventhisdefinitioni, we can analyzethe"remarriage" choiceof
separatedwomen.Welfarereceiptis definedas receiving a dollaror moreof incomefrom
AFDC or the "otherwelfare"category used in thePSID. 12
We are interestedin analyzingthe determinanits of the trichotomous choice of
remarriage, welfarereceipt,and remaining singlewithoutwelfarereceipt.The motivation
includesunderstanding the potentialroleofAFDC incomein discouraging remarriage as
wellas themoregeneraldeterminants of remarriage decisions.
One explanation, castin termsofindividualcharacteristics, mightfocuson suchthings
as a woman'sage and education,thenumberofchildrenshe has, and whether she resides
in an urbanarea. We estimatethismodelwiththeMNLGT model.
A different explanationmightconsider,instead,the exogenousincome(the income
availableto a womanat zerohoursofwork)and hernet(after-tax) wageratein each ofthe
threealternatives. This correspondsto a modellikeequation(6) in whichthevalue of an
alternativeis a functionof itscharacteristics, hereexogenousincomeand prices.Techni-
cally,we are usingtheconceptof indirectutility functions in whichthemaximumutility
(satisfaction)
availableto an individualin an alternative dependson itsexogenousincome
and thesetofprices(in our model,thewagerate)it provides.
Consider,first, theexogenousincomeavailableto a womanin each alternative. While
someincome,suchas childsupport, and incomefromdividends and interestareunaffected
by her choice of alternative, othercomponenits varysystematically by alternatives. For
instance,ifshe wereto acceptwelfare, shewouldreceivethelegallymandatedbenefits paid
in herstateofresidence,givenherfamilysize and otherincome.Ifshe wereto marry, she
wouldbe ineligibleforwelfareand benefits, 13 butshe wouldhaveaccessto someportion of
her new husband'sincome. If she remainedsinglewithoutacceptingwelfare,she would
receivealimonyand/orchildsupportincome,ifany,plusanyincomefromdividendsand
interest.
Her after-tax wage ratewouldalso differ acrossalternatives,even thoughthe market
422 Demography,
Vol. 25, No. 3, August1988
Table2. Characteristics
ofPSID Sampleof
WhiteWomenUndergoing Divorce
orSeparation,
1969-1982
Samplesize
Persons 460
Person-years 1,269
alternatives
Person-year 3,304
Married 1,269
welfare
Single/no 1,269
Single/welfare 766
Individual
characteristics
Age 30.9
Years ofeducation 12.1
No. of children 1.4
Urban residence 0.28
Economiccharacteristics
ofthealternatives
($)
AFDC income
(thousands) 3.68
Spouse incomea
(thousands) 16.26
Wage ratea
Married 4.80
Single/nowelfare 5.59
Single/welfare 1.30
Table3. EstimatesofRemarriage
and WelfareChoicesofDivorcedand SeparatedWhiteWomen,
PSID, 1969-1982
Summary
This articlehas providedan introductionto and illustration
ofthe use of conditional
logitto estimatemultiple-categorydiscrete-choice
problems.CLGT is closelyrelatedto the
better-known MNLGT model,but it derivesfromdifferent behavioralassumptions and is
estimatedin different form.The CLOT modelis appropriate wheneverit is reasonableto
Multinomialand ConditionalLogitModels 425
Notes
A reviewof 10 issues of Demography (publishedbetweenFebruary1984 and May 1986)
produced10 examplesof discrete-choice researchusinga logitmodel. Seven of the 10 involved
two-categorydependent variables-Masseyand Mullen(1984)analyzedthepresenceofyoungchildren
in a household,Landale and Guest (1985) mobilityplans and actions,Tienda and Glass (1985)
women'slabor forceparticipation, Entwisleand her colleagues(Entwisleet al., 1984; Entwisle,
Mason, and Hermalin,1986)contraception behavior,DaVanzo and Habicht(1986) infantmortality,
and Bellerand Graham(1986) the presenceof a child-support award.Examplesof three-category
choice models include Lehrerand Kawasaki's(1985) analysisof a child care modal choice and
Leibowitz,Eisen, and Chow's (1986) analysisof teenagepregnancy decisionmaking.Robinsand
Dickinson(1985) estimated a four-categorymodelofwelfareand childsupport.
2 The threemultiple-category modelsidentified in note 1 are all examplesofmultinomial logit.
Leibowitz,Eisen, and Chow (1986) used conditionallogitto describewhat is more commonly
consideredmultinomial logit.
3 Although noneofthestatistical methodsdescribedhereis new,and discussions ofsomeofthe
issuescan be foundin statisticaland econometrics texts(see Ben-Akiva and Lerman,1985;judgeetal.,
1980;Maddala, 1983),we knowofno applieddiscussionthatfocusesexplicitly on theissuesdiscussed
here.
4 For modelingand estimation purposes,thedistinction drawnbetweenMNLGT and CLGT is
usefuland instructive.The modelsdo, however,sharea commonlikelihoodfunction; see thethird
sectionfora discussionofthis.
5As in CLGT, themixedlogitmodelusesthealternative as theunitofanalysis.Whatwe call
mixedlogitis sometimes referredtoas "multinomial logit,"withthepureMNLGT and CLGT models
treatedas specialcases in whichonlythecharacteristics ofalternatives or ofindividualsare used (see
Amemiya,1985;Ben Akivaand Lerman,1985).We findthisterminology confusing,sinceitsuggests
model in questionis the morefamiliarand significanitly
thatthe statistical pure MNLGT
different
modelof equation(1).
426 Vol. 25, No. 3, August1988
Demography,
Acknowledgments
This articlcis basedon researclh
supportedbv NationalInstitute and Human Developmeicnit
forChild I-Health
GrantIROl HD 19339-01.It has beniefited fromhelpfulcommcnits bv Johln
Bounid,DorothvDuncani,Robert
Hutchens,WillardRodgers,GarvSolon, and ArlandThornitoni. We also thaniktwoanonivmousrefereesand the
deputveditorforthcircommncits.h'lec orderoftheauthors'namtes was determinied
randomlv.
References
Allison,P. D. 1982. Discretetimemetlhods fortheanialvsisofeventhistories.Pp. 64-98 in S. Lclhnhlardt(ed.),
SociologicalMethodology. San Franicisco:Jossev-Bass.
1984. EventHistoryAnalysis.BeverlyHlills,Calif.:Sage.
Amemiva,T. 1985. AdvancedEconometrics. Cambridge,Mass.: HarvardUniversity Press.
Beni-Akiva,M., and S. R. Lermani.1985. DiscreteChoiceAnalysis.Cambridgc,Mass : MllT Press.
Beller,A. H., and J.W. Gralham.1986. Childsupport awards:Differenitials
and trenids
byraceand maritalstatus.
Demography 23:231-246.
DaVanzo, J.,and J.-P.Habicht.1986. Infanit mortalitvdcclinein Malaysia,1946-1975:Tlc rolesofchanigcs in
variablesand changesin thestructure ofrelationislhips.
Demography 23:143-160.
Entwisle,B., A. I. Hermalin,P. Kamnuaiisilpa,and A. Chamratrithironig. 1984. A multilevelmodelof familv
planningavailabilitvand contraceptiveuse in ruralTlhalaind.Demography 21:359-574.
Eiutwisle,B., W. M. Mason, and A. I. Hermalin.1986. The multileveldepenidcnce of contraceptive
use on
sociocconiomic development and familyplanninig program strenigtlh.
Demography 23:199-216.
Hausman,J.A., and D. McFadden. 1984. Specificationi testsforthemultinomial logitmodel.Econometrica 52:
1219-1240.
Hoffman, S. D., and G. J.Duncan. 1987. Remarriage and WelfareChoicesofDixvorced and SeparatedWomeni.
Unpublishedmanuscript, UniversitvofDelaware,Dept. of Economics.
Judge,C., W. Griffins, R. Carter,and T. Lee. 1980. Tlhe Tlheoryand PracticeofEconometrics. Ncw York:Wilex.
Landale, N. S., and A. Guest. 1985. Constraints,satisfaction and rcsidcntialnmobilitv: Spearc's model
reconsidered. Demography 22:199-222.
Lee, L. 1983. Generalizedeconomctric modelswithsclectivity. Econometrica 51:507-512.
Lehrer,E., and S. Kawasaki.1985. Child carearrangeeniits An analsis oftwo-earner
and fertility: households.
Demography 22:499-514.
Leibowitz,A., M. Eiscen,and W. K. Chow. 1986. An economicmodclofteenagepregnancy decisioni-makinig.
Demography 23:67-78.
Maddala, G. S. 1983. Limited-Dependent and Qualitative Variables in Econometrics.Cambridge,U.K.:
CambridgeUniversity Press.
Massey, D. S., and B. P. Mullan. 1984. A demonstrationi of the effectof seasonialmigrationion fertility.
Demography 21:501-518
McFadden,D. 1973. Conditionial logitanalvsisofqualitativechoicebehavior.Pp. 105-135in P. Zaremiibka (ed.),
Frontiersin Econometrics.New York:Wiley.
1981. Econometric nmodelsofprobabilistic
choice.Pp. 198-272in C. F. Maanski and D. McFaddeni(eds.),
Structural AnalysisofDiscreteData WithEconometric Applications. Cambridge,MNass.. mIT Press.
Robins,P. K., and K. P. Dickinson.1985. Clild supportand welfare depenidcnce: A Multinomiallogitanalvsis.
Demography 22:367-380.
Tienda, M., and J.Glass. 1985. Houselhold structure
and laborforceparticipationi ofblack,Hispanic,and white
workers. Demography 22:381-394.